This paper focuses on the automatic diagnosis of faults in new energy vehicles (NEVs) using sensor data, addressing the broader problem of predictive maintenance in electric drivetrain systems. This paper will also determine whether machine learning (ML) models can reliably classify NEV sensor readings into normal operation and multiple distinct fault categories, and to make those models' predictions interpretable rather than opaque, so that their outputs can be trusted and acted upon by maintenance engineers. This was tackled by training and comparing four classical ML classifiers, Logistic Regression, Naive Bayes, Random Forest, and XGBoost, on the NEV Fault Diagnosis Dataset, evaluating each on a stratified 50:50 train/test split using macro F1 score, 5-fold cross-validation, and linear weighted Cohen's Kappa. The best-performing model, Random Forest, was then analyzed using SHAP (SHapley Additive exPlanations) to identify which sensor features drove its predictions, both globally across the dataset and locally for individual predictions, and this analysis was cross-checked against a correlation matrix computed directly from the raw sensor data. This paper is organized as follows: Section I introduces the problem of NEV fault diagnosis and the role of explainable machine learning; Section II reviews related work in machine learning-based fault diagnosis; Section III describes the dataset, preprocessing, and model training methodology; Section IV presents the quantitative results of the four models and a qualitative SHAP-based analysis of the best-performing model; Section V discusses the key findings, the best-performing model, and the limitations of this work; and Section VI presents the conclusion.

Keywords: Machine Learning; Fault Diagnosis; Electric Vehicles; Random Forest; SHAP